Synchronizer for a fault tolerant multiple node processing system

• US 4,816,989 A
• Filed: 04/15/1987
• Issued: 03/28/1989
• Est. Priority Date: 04/15/1987
• Status: Expired due to Term

• Current Assignee: Allied Corporation Columbia Road and Park Avenue Morris Township New Jersey
• Original Assignee: Alliedsignal Inc. (Honeywell International Inc.)
• Inventors: Finn, Alan M., Walter, Chris J., Kieckhafer, Roger M.
• Primary Examiner(s): Zache, Raulfe B.
• Application Number:
  US07/038,818
• Time in Patent Office: 713 Days
• Field of Search: 364/200, 364/900
• US Class Current: 709/248
• CPC Class Codes: G06F 11/0724 in a multiprocessor or a multi-core unit multiprocessors per se G06F15/80
• CPC Class Codes: G06F 11/076 by exceeding a count or rate limit, e.g. word- or bit count limit
• CPC Class Codes: G06F 11/10 Adding special bits or symbols to the coded information, e.g. parity check, casting out 9's or 11's
• CPC Class Codes: G06F 11/1425 by reconfiguration of node membership
• CPC Class Codes: G06F 11/1482 by means of middleware or OS functionality
• CPC Class Codes: G06F 11/1658 Data re-synchronization of a redundant component, or initial sync of replacement, additional or spare unit
• CPC Class Codes: G06F 11/181 Eliminating the failing redundant component
• CPC Class Codes: G06F 11/182 based on mutual exchange of the output between redundant processing components
• CPC Class Codes: G06F 11/187 Voting techniques
• CPC Class Codes: G06F 11/188 where exact match is not required
• CPC Class Codes: G06F 11/202 where processing functionality is redundant redundant communication control functionality G06F11/2005, redundant storage control functionality G06F11/2089
• CPC Class Codes: G06F 15/161 Computing infrastructure, e.g. computer clusters, blade chassis or hardware partitioning casings, cabinets, racks or drawers for data centers H05K5/00
• CPC Class Codes: G06F 9/4881 Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues

Abstract

A synchronizer for each node in a multiple node processing system having a message interface for receiving sync and pre-sync time-dependent message, counter means for generating a local time, a time stamp memory having an entry for each node in the multiple node processing system, a time stamper responsive to receiving a time-dependent message from a node for storing the local time in the entry of said time stamp memory for that node to generate a time stamp. The synchronizer has a time stamp voter for generating a medial time stamp value from all the time stamps in the time stamp memory, a sync correction generator for generating a sync delta having a value corresponding to the difference between the voted time stamp and the node'"'"'s own time stamp, means for adding said sync delta to a nominal transmission timing interval to generate an actual transmission timing interval, and a message generator for generating a pre-sync time dependent message passed to the transmitter at the end of the nominal transmission timing interval and for generating a sync time-dependent message passed to the transmitter at the end of said actual transmission timing interval.

Claims

1. In a multiple node processing system having a plurality of nodes, an application processor for processing a predetermined set of tasks and an operations controller for controlling its own node, the operations controller having a transmitter for transmitting a synchronizer for synchronizing its own node with all of the nodes in the system through an exchange of inter-node time-dependent messages comprising alternating presync and sync time-dependent messages separated in time by a nominal timing interval, said synchronizer comprising:

a message interface for receiving the sync and pre-sync time-dependent messages;

counter means for generating a local time;

a time stamp memory having one entry for each node in the system, each entry storing a time stamp for a most recent time-dependent message received from an associated node;

a time stamper connected to said message interface, said counter means and said time stamp memory, said time stamper responsive to receiving a time-dependent message from each node for generating a time stamp corresponding to a local time at which said time-dependent message is received and for storing said time stamp in said time stamp memory in an entry associated with the node from which the time-dependent message was received;

a time stamp voter connected to said time stamp memory for generating a voted time stamp corresponding to a medial value of said time stamps stored in said time stamp memory for said pre-sync time-dependent messages;

a sync correction generator connected to said time stamp memory and said time stamp voter for generating sync delta having a value corresponding to a difference between said voted time stamp and the time stamp of its own pre-sync time-dependent message;

means connected to sid sync correction generator for adding said sync delta to a nominal transmission timing interval for generating an actual transmission timing interval, said nominal transmission timing interval corresponding to a nominal timing interval between the end of the transmission of said sync time-dependent message and a passing of said pre-sync time-dependent message to said transmitter; and

message generator means connected to said means for adding for generating said sync and pre-sync time-dependent messages, said message generator means passing said pre-sync time-dependent messages to said transmitter at an end of said nominal transmission timing interval and passing said sync time-dependent messages to said transmitter at an end said actual transmission timing interval.

2. The synchronizer of claim 1 having a timing signal generator for generating at least a first timing signal coincident with an end of the transmission of said sync and pre-sync time-dependent messages and a hard error window signal defining an interval of time bracketing said first timing signal during which all of said sync and pre-sync time-dependent messages should be received from the nodes which are in synchronization with its own node, said synchroziner having a hard error window detector connected to said message checker interface, said timing signal generator and said time stamper for inhibiting said time stamper from recording a time stamp in said time stamp memory for each node whose time-dependent message was not received within said hard error window, and for recording a hard error for each node whose time-dependent message was not received within said hard error window.

3. The synchronizer of claim 2 wherein said timing signal generator generates clock pulses and a second timing signal which divides a first nominal timing interval into a plurality of subintervals, said counter means comprising:

a first counter means connected to said timing signal generator for counting each occurrence of said second timing signal during said first timing signal to generate a sub-interval count;

a second counter means connected to said timing signal generator for counting said clock pulses between each occurrence of said second timing signal to generate a clock count; and

wherein said time stamp comprises a coarse time stamp portion containing said sub-interval count and a fine stamp portion containing said clock count.

4. The synchronizer of claim 3 wherein said time stamp voter generates a voted coarse time stamp corresponding to a medial value of said sub-interval count stored in said time stamp memory and a voted fine time stamp corresponding to a medial value of said clock count stored in said time stamp memory and wherein said sync correction generator generates a coarse delta equal to a difference between said voted coarse time stamp and its own coarse time stamp and a fine delta equal to a difference between said voted fine time stamp and a number of clock counts in its own time stamp.

5. The synchronizer of claim 4 wherein said sync and pre-sync time-dependent messages have a message type code, said message type code of said sync time-dependent message is distinguished from said message type code of said pre-sync time-dependent message by function bits contained therein, said synchronizer further has an expected message detector connected to said message checker interface and said time stamper for recording a sequence error for each node which sends a message within the hard error window which has a wrong message type code and records a sequence error for each node in which said sync and pre-sync time-dependent messages are not received in an alternating manner, said expected message detector inhibiting said time stamper from recording a time stamp in said time stamp memory for each node which sent a message having a detected error sequence.

6. The synchronizer of claim 5 having:

an in-sync-with register having a flag bit for each node in the processing system; and

an in-sync-with detector connected to said hard error window detector and said in-sync-with register for setting the flag bit in said in-sync-with register for each node whose time stamp differed from the time stamp stored for its own node by a time interval less than said hard error window to generate an in-sync-with vector.

7. The synchronizer of claim 6 wherein said message generator includes said in-sync-with vector in at least said sync time-dependent message.

8. The synchronizer of claim 7 wherein said synchronizer has at least three modes of operation, a cold start mode initiated when power to the multiple node processing system is first turned on, a warm start mode in which an operating set of nodes are in synchronization with each other but its own node is not in snychronization with said operating set, and an operating mode in which its own node is included in the operating set, said synchronizer has:

an operating flag connected to said message generator, said operating flag when set to true signifying the synchronizer is in said operating mode of operation;

a warm start flag connected to said message generator, said warm start flag when set to true signifying the synchronizer is in said warm start mode of operation;

a scratch pad memory connected to said means for adding and said message generator, said scratch pad memory having an entry for each node in the processing system, each entry storing the in-sync-with vector received from the node associated with the entry in said sync time-dependent messages to form an in-sync-with matrix; and

wherein said message generator will include said in-sync-with matrix in said pre-sync time-dependent messages in response to said operating flag and said warm start flag being set to false.

9. The snychronizer of claim 8 having:

a Byzantine memory connected to said message checker interface for storing said in-sync-with matrix received from each node to generate a three-dimensional in-sync-with matrix cube in response to said operating flag and warm start flag being set to false;

a Byzantine voter connected to said Byzantine memory for generating a Byzantine voted in-sync-with vector from said in-sync-with matrix cube; and

operating condition detector means connected to said Byzantine voter, said operating flag and said warm start flag for detecting an existence of an operating set of nodes when said Byzantine voted in-sync-with vector signifies that more than a minimum number of nodes are in synchronization with each other, to set said operating flag to true when an operating set of nodes exists and its own node is included in said operating set, and to set said warm start flag to true when an operating set of nodes exists and its own node is not included in said operating set of nodes.

10. The synchronizer of claim 9 wherein the operations controller includes a fault tolerator subsystem, said operating condition detector means passes said Byzantine voted in-sync-with vector to said fault tolerator subsystem as an initial system state for its own node.

11. The synchronizer of claim 10 wherein said hard error window detector enables said time stamper to time stamp all pre-sync time-dependent messages received from the nodes for a predetermined listening period, larger than said hard error window, in response to said operating flag being set to false.

12. The synchronizer of claim 11 wherein said predetermined listening period is equal to the nominal timing interval said interval of time of said hard error window.

13. The synchronizer of claim 12 wherein said time stamp entry includes a sync flag bit signifying the stored time stamp is for a sync time-dependent message, a pre-sync flag bit signifying the stored time stamp is for a pre-sync time-dependent message and an updated flag when set to true signifying that a new time stamp was recorded for that node during a current time stamp interval, said synchronizer further has:

a sync message counter connected to said message checker interface for counting a number of sync time-dependent messages received during the current time stamp interval to generate a sync message count;

a pre-sync counter connected to said message checker interface for counting a number of pre-sync time-dependent messages received during the current time stamp interval to generate a pre-sync message count; and

a comparator connected to said sync and pre-sync message counters for generating a first signal when said sync message count is greater than said pre-sync message count, a second signal when said pre-sync message count is greater than said sync message count, and a third signal when said sync and pre-sync message counts are equal;

wherein said time stamp voter uses a time stamp having said updated flag and said sync flag set to true, to generate said voted time stamp value in response to said first signal and uses a time stamp having said updated flag and said pre-sync flag set to true to generate said voted value in response to said second signal; and

wherein said sync correction generator sets said coarse delta and said fine delta equal to zero in response to said third signal generated by said comparator.

14. The synchronizer of claim 4 wherein said sync correction generator has means responsive to said coarse delta being greater than zero for generating a substitute coarse delta having a value equal to said coarse delta minus one, and a substitute fine delta having a value equal to a predetermined maximum fine delta value, and responsive to said coarse delta being less than zero for generating said substitute coarse delta having a value equal to said coarse delta plus one and a substitute fine delta having the 2'"'"'s complement of said predetermined maximum fine delta value.

15. The synchronizer of claim 9 wherein the operations controller for each node has a scheduler for selecting the tasks to be executed by the applications processor, and each scheduler has a scheduling table from which the tasks are selected and wherein each node transmits a task interactive consistency message to every node in the processing system and wherein each task interactive consistency message includes a task completed vector identifying each node which reported to have completed a task and a branch condition bit which is used by said scheduler to identify a successor task to the completed task, said synchronizer has a message memory connected to said message checker interface and said Byzantine voter memory, said message memory having an entry for each node in which said task completed vector and said branch condition bit are stored and wherein said Byzantine voter generates a Byzantine voted task completed vector and Byzantine voted branch condition bits which are passed to said scheduler for updating its scheduling table.

16. The synchronizer of claim 5 wherein said operations controller includes a fault tolerator and said timing signal generator generates a soft error window defining an interval of time bracketing said first and second timing signals which has a duration less than said hard error window and wherein said synchronizer includes:

a soft error window detector for generating a soft error for each node whose time-dependent message is received within said hard error window and outside of said soft error window; and

an error reported for reporting said hard errors, soft errors, and said sequence errors to said fault tolerator.

17. The synchronizer of claim 15 wherein said message generator means generates said task interactive consistency messages when said synchronizer is in said operating mode and passes them to the transmitter at a predetermined time prior to their transmission, said transmitter including means to start the transmission of each task interactive consistency message such that an end of its transmission is coincident with said second timing signal generated by said timing signal generator.

18. The synchronizer of claim 17 wherein said timing signal generator generates a last sub-interval signal identifying a beginning of a last sub-interval in each nominal timing interval, and wherein each second timing signal is also bracketed by said hard error window signal, said actual and nominal transmission timing intervals used by said message generator to pass said sync and pre-sync messages to said transmitter, when said synchronizer is in said operating mode, is measured from an end of said hard error window bracketing the second timing signal coincident with a last sub-interval signal and wherein said task interactive consistency message is passed to said transmitter at the end of said nominal transmission timing interval measured from said end of the hard error window bracketing a preceding second timing signal.

19. The synchronizer of claim 15 wherein said Byzantine voter comprises:

a data register having a plurality of bit positions, one bit position for each node in said processing system;

an address generator connected to said Byzantine memory for generating addresses enabling said Byzantine memory to unload its contents into said data register in a selectable sequence;

a plurality of AND gates, each having a first input connected to a respective one of said plurality of bit positions in said data register, each of said AND gates further having a second input and an output;

decoder means responsive to said addresses generated by said address generator for generating individual signals applied to said second input of said plurality of AND gates, said decoder means generating an inhibit signal which deactivates one of said plurality of AND gates in a sequential manner each time new data ia loaded into said data register;

a plurality of multiplexers, one associated with each AND gate, each of said plurality of multiplexers having a first input connected to said output of its associated AND gate, a second input, and an output;

a message counter for generating a 2'"'"'s complement reference number which is applied to said second input of each of said plurality of multiplexers, said 2'"'"'s complement reference number corresponding to the 2'"'"'s complement of one half of a number of messages whose contents are stored in said Byzantine memory;

a plurality of adders for generating sum data and voted data, each adder having a first input connected to said output of a respective one of said multiplexers, a second input, a sum output outputting said sum data and an overflow output outputting said voted data;

a plurality of accumulators for storing said sum data, each accumulator having an input connected to said sum output of a respective one of said plurality of adders, each of said plurality of accumulators having an output connected to said second input of said adder whose sum output is connected to its input to return said sum data to said adder;

a voted data register having a bit position associated with a respective one of said plurality of adders, each bit position connected to and storing said voted data generated by its associated adder;

a 3;

1 output multiplexer having a first state passing said voted data stored in said voted data register to the Byzantine memory where it is temporarily stored, a second state passing said voted data to said operating condition detector and a third state passing said voted data to the scheduler; and

control logic means responsive to a message type code of the data being processed for generating sequence signals identifying the sequence in which said Byzantine memory is to be addressed to load said data register, for generating a load accumulator signal enabling said plurality of multiplexers to load said 2'"'"'s complement reference number into said accumulators through said plurality of adders, for generating an add data signal enabling said AND gates each time said data register is loaded with new data, for generating a first multiplexer signal placing said output multiplexer in said first state, a second multiplexer signal placing said output multiplexer in said second state and a third multiplexer signal placing said output mulitplexer in said third state.

20. The synchronizer of claim 19 wherein said control logic means is responsive to a message type code for said pre-sync time-dependent message to generate a first sequence signal causing said address generator to generate addresses loading said data register with in-sync-data bytes from said cube matrix row by row from a front row of the cube to a back row of the cube, and to generate said first multiplexer signal to store said voted data stored in said voted data register in said Byzantine memory to form a two-dimensional in-sync-with matrix, said control logic means further being responsive to a loading of a last voted in-sync-with vectors into said Byzantine memory to generate a second sequence signal causing said address generator to load each column of said two-dimensional in-sync-with matrix, column by column into said data register and to generate said second multiplexer signal enabling said output multiplexer to pass a final content of said voted data register to said operating condition detector.

21. The synchronizer of claim 20 wherein said control logic means is responsive to a message type code of said task interactive message indicating the data stored in the Byzantine memory is a two-dimensional task completed vector matrix and a two-dimensional branch condition bit matrix to generate said second sequence signal causing said address generator to generate addresses loading said data register with the data from each column of said task completed vector matrix, column by column, and to generate said third multiplexer signal to pass voted data stored in said voted register to said scheduler after each column of task completed vectors has been processed and wherein said second sequence signal further causes said address generator to generate addresses loading said data register with the data from each column of said branch condition bit matrix, column by column, and to generate, a second time, said third multiplexer signal to pass said voted data to said scheduler after each column of said branch condition bit matrix has been processed.

22. A Byzantine voter comprising:

a voter memory having transverse and lateral rows and columns for storing a three-dimensional cube of data, said transverse and lateral rows and columns each having an equal number of bit positions;

a data register connected to said vector memory having a plurality of bit positions, said plurality of bit positions at least equal to said number of bit positions in said transverse and lateral rows and columns of said voter memory;

an address generator connected to said voter memory for generating addresses enabling said voter memory to unload said data stored in said rows and columns into said data register, one row or one column at a time in a selected sequence;

a plurality of AND gates each having one input connected to a respective one of said bit positions in said data register, said AND gates each having a second input and an output;

decoder means connected to said address generator and said AND gates responsive to said addresses generated by said address generator for generating individual signals applied to said second input of each of said plurality of AND gates, said decoder means generating an inhibit signal which deactivates one of said plurality of AND gates in a sequential manner each time new data is loaded into said data register;

a plurality of multiplexers, one associated with each AND gate, each of said plurality of multiplexers having a first input connected to said output of its associated AND gate, each multiplexer further having a second input and an output;

a message counter connected to said plurality of multiplexers for generating a reference number which is applied to said second input of each of said plurality of multiplexers, said reference number corresponding to a 2'"'"'s complement of one half of the number of data values to be processed;

a plurality of adders for generating sum data and voted data, each having a first input connected to said output of a respective one of said multiplexers, each of said adders having a second input, a sum output outputting said sum data and an overflow output outputting said voted data;

a plurality of accumulators for storing said sum data, each accumulator having an input connected to said sum output of a respective one of said plurality of adders, each of said plurality of accumulators having an output connected back to said second input of the adder to add said sum data back into said adder;

a voted data register having a plurality of bit positions, each bit position associated with a respective one of said plurality of adders, each of said bit positions connected to the overflow output of its associated adder and storing said voted data;

an output gate connected to said voted data register for passing said voted data stored in said voted data register back to said voter memory in response to a recycle signal; and

control logic means responsive to an input signal signifying a particular type of data stored in said voter memory for generating sequence signals identifying the sequence that data from said voter memory is to be loaded into said data register, for generating a load reference number signal enabling said plurality of multiplexers to load said reference number into said plurality of accumulators through said adders, for generating an add signal enabling said multiplexers to pass said data stored in said data registers to said adders to be added to said sum data stored in said accumulators, and for generating said recycle signal when necessary enabling said voted data to be stored in said voter memory.

23. The Byzantine voter of claim 22 wherein said control logic means is responsive to an input signal identifying the data being stored in said voter memory is a plurality of data matrices forming a cube matrix having three axes, said control logic means having means for generating a first sequence signal causing said address generator to generate addresses along one of said three axes of said cube to sequentially load said data register, one byte at a time, until all of the data in the cube is processed a first time and for generating said recycle signal storing said voted data back in said voter memory to form a two-dimensional matrix of voted data, said control logic means further being responsive to loading of said voted data from said voted data register into said voter memory for generating a second sequence signal causing said address generator to generate addresses along a different axis of the cube to sequentially load said data register, one byte at a time, until all the data in said two-dimensional data matrix of voted data is processed and a final Byzantine voted data value is stored in said voted data memory.

24. The Byzantine voter of claim 23 wherein said control logic means is responsive to an input signal identifying the voted data stored in said data memory is a two-dimensional data matrix, said control logic means generates said second sequence signal causing said address generator to generate addresses along said different axis to generate said Byzantine voted data value.